IC-110Phi-3 (3.8B) and Llama-3 (8B) with triplet prompting outperform GPT-4 with pairwise prompting on causal graph orientation

Aniket Vashishtha, Abbavaram Gowtham Reddy, Abhinav Kumar, Saketh Bachu, Vineeth N. Balasubramanian, Amit Sharma

SourceCausal Order: The Key to Leveraging Imperfect Experts in Causal Inference

The paper evaluates whether smaller open-weight models can match or exceed a larger proprietary model on causal reasoning. Using the triplet method (which queries three variables at a time and enforces acyclicity), Phi-3 and Llama-3 8B produce cycle-free causal orders with lower SHD than GPT-4 using standard pairwise prompting. On the 20-node child dataset, Phi-3 triplet achieves SHD 69 and Llama-3 triplet achieves SHD 129, compared to GPT-4 pairwise's SHD 148, while both smaller models produce zero cycles versus GPT-4's >>10k cycles.

Evidence
correlational
Key metric
child dataset: GPT-4 pairwise SHD 148, >>10k cycles; Phi-3 triplet SHD 69, 0 cycles; Llama-3 triplet SHD 129, 0 cycles
Caveat
The comparison is confounded by the prompting strategy: the smaller models use the triplet method while GPT-4 uses pairwise. The paper does not report GPT-4 with triplet prompting on the same datasets for a controlled comparison (Table A10 shows GPT-4 triplet on a subset but with different metrics).
Model
Phi-3, Llama 3, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Concepts
Scale-dependent behaviour
Datasets
bnlearn [eval]
Related work
Jiralerspong et al. 2024 [compared-to]
Related findings
IC-109
Extraction
automatic-extraction